Open Source vs Closed Source AI Models for Financial Analysis in 2026
Closed source models still score higher on financial analysis benchmarks in November 2026, and the difference to open weight models is now 2.5 points on Vals AI Finance Agent v2.1. Claude Opus 5.5 scores 67.9% at $8.91 per task, and DeepSeek V4.1 Flash, the highest scoring open weight model, scores 65.4% at $0.37 per task. Open weight models score higher than every closed source model in three of nine task categories: Earnings Analysis, Precedents and Comparables. This blogpost compares closed and open weight models on financial analysis and sets the results in relation to cost and speed.
On this page:
- Accuracy on finance
- Finance Agent v2.1 ranking
- Open vs closed models as groups
- Open weight models released since September
- Excel modeling, tax and the Vals Finance Index
- Results by task
- Categories led by open weight models
- Largest gaps
- Cost and speed
- Cost per task
- Time per task
- Data control and licensing
- FAQ
- Related reading
Accuracy on finance
Finance Agent v2.1 ranking
Finance Agent v2.1 from Vals AI tests models on analyst questions based on SEC filings. It replaced Finance Agent v2 in November 2026 with corrected questions and rubrics, a new jury of three judge models and a rule that fails answers which give two or more candidate values for one item. Each model uses six tools: EDGAR search, web search, an HTML page parser, a retrieval tool, a calculator and a price history tool. The scores below are on the private test set of 449 questions, as published on 5 November 2026. Cost per task is the API cost of one question on the provider Vals AI used for the run. Time per task is the average time from question to final answer.
| Rank | Model | Developer | Type | Accuracy | Cost per task | Time per task |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | Closed source | 67.9% | $8.91 | 33.0 min |
| 2 | Gemini 4 Argon | Closed source | 67.7% | $4.32 | 13.2 min | |
| 3 | Muse Wonder Full | Meta | Closed source | 67.6% | $2.68 | 14.2 min |
| 4 | Inkling 2 Small | Thinking Machines | Weights announced | 67.5% | $0.18 | 4.9 min |
| 5 | Claude Sonnet 5.5 | Anthropic | Closed source | 67.3% | $6.95 | 34.4 min |
| 6 | Gemini 4 Carbon | Closed source | 67.2% | $7.29 | 29.7 min | |
| 7 | Gemini 3.8 Flash | Closed source | 65.8% | $1.99 | 7.4 min | |
| 8 | Muse Wonder | Meta | Closed source | 65.6% | $2.62 | 5.2 min |
| 9 | DeepSeek V4.1 Flash | DeepSeek | Open weight | 65.4% | $0.37 | 16.0 min |
| 10 | Claude Haiku 5.5 | Anthropic | Closed source | 65.4% | $1.64 | 31.1 min |
| 11 | GLM 5.3 | Z.ai | Open weight | 65.0% | $1.03 | 20.1 min |
| 12 | Hy4 | Tencent | Weights announced | 64.7% | $1.28 | 37.3 min |
| 13 | MiMo V2.6 Pro | Xiaomi | Open weight | 64.3% | $0.18 | 11.4 min |
| 14 | MiMo V2.6 Flash | Xiaomi | Open weight | 63.3% | $0.07 | 15.6 min |
| 15 | DeepSeek V4.1 Flash 1017 | DeepSeek | Open weight | 62.9% | $0.47 | 15.2 min |
| 16 | GPT 6.1 Sol | OpenAI | Closed source | 61.8% | $1.63 | 19.1 min |
| 17 | Mistral Large 4 | Mistral AI | Open weight | 61.7% | $0.35 | 6.5 min |
| 18 | Hy4 Preview | Tencent | Open weight | 61.6% | $0.67 | 25.0 min |
| 19 | Kimi K3.1 Pro | Moonshot AI | Open weight | 61.5% | $2.29 | 35.0 min |
| 20 | Mistral Large 4 Preview | Mistral AI | API preview | 61.4% | $0.52 | 16.4 min |
The top 20 models include eight open weight models: DeepSeek V4.1 Flash in 9th place, GLM 5.3 in 11th, MiMo V2.6 Pro in 13th, MiMo V2.6 Flash in 14th, DeepSeek V4.1 Flash 1017 in 15th, Mistral Large 4 in 17th, Hy4 Preview in 18th and Kimi K3.1 Pro in 19th. Muse Spark Open follows in 21st place at 60.7%, GLM 5.3 Flash in 22nd at 60.6% and Kimi K3 in 25th at 60.0%. Inkling 2 Small from Thinking Machines scores 67.5% in 4th place at $0.18 per task. Thinking Machines released it on 5 November 2026 through its Tinker platform and announced the weights under Apache 2.0 for the following week. Tencent released the final Hy4 on 4 November 2026 as an API model and announced the weights for a later date; the weights of Hy4 Preview from August are published under Apache 2.0. The full ranking of open weight models with parameters, licenses and hardware requirements is in Best open source models for financial analysis.
Open vs closed models as groups
The group figures cover 33 of the 37 models on the leaderboard: 19 open weight models, whose weights are published on Hugging Face, and 14 closed source models, which are available through an API only. Four models are left out of the group figures: Inkling 2 Small and Hy4, whose weights were announced but unpublished on 6 November 2026, Mistral Large 4 Preview, the API build that preceded the published Mistral Large 4 weights, and Qwen 3.8 Max 0902, whose weight availability is unconfirmed.
| Measure | Open weight models | Closed source models |
|---|---|---|
| Models on the leaderboard | 19 | 14 |
| Highest accuracy | 65.4% (DeepSeek V4.1 Flash) | 67.9% (Claude Opus 5.5) |
| Average accuracy of the top 5 | 64.2% | 67.5% |
| Median accuracy of all models in the group | 60.6% | 65.5% |
| Models at 60% accuracy or higher | 11 | 11 |
| Median cost per task of the top 5 | $0.37 | $6.95 |
| Median time per task of the top 5 | 15.6 min | 29.7 min |
The top 5 open weight models are DeepSeek V4.1 Flash, GLM 5.3, MiMo V2.6 Pro, MiMo V2.6 Flash and DeepSeek V4.1 Flash 1017. The top 5 closed source models are Claude Opus 5.5, Gemini 4 Argon, Muse Wonder Full, Claude Sonnet 5.5 and Gemini 4 Carbon.
Open weight models released since September
Eight open weight models on the leaderboard had their weights published between 15 October and 4 November 2026.
| Model | Developer | Weights published | License | Finance Agent v2.1 |
|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | DeepSeek | 17 Oct 2026 | MIT | 62.9% |
| Mistral Large 4 | Mistral AI | 29 Oct 2026 | Mistral Ecosystem Open Weight License | 61.7% |
| Kimi K3.1 Pro | Moonshot AI | 4 Nov 2026 | Kimi K3.1 Pro License | 61.5% |
| Muse Spark Open | Meta | 15 Oct 2026 | Muse Model License | 60.7% |
| Step 5 Preview | StepFun | 15 Oct 2026 | StepFun Model License 1.0 | 58.9% |
| MiniMax M3.1 Flash | MiniMax | 23 Oct 2026 | MiniMax Community License | 58.8% |
| Beam 1.0 | Reflection AI | 16 Oct 2026 | Apache 2.0 | 51.8% |
| Nemotron 3.5 Super VL | NVIDIA | 15 Oct 2026 | OpenMDW 1.1 | 47.7% |
Muse Spark Open is the open weight version of Meta's Muse Spark and runs under a custom license with commercial conditions. Mistral Large 4, Beam 1.0, Muse Spark Open and Nemotron 3.5 Super VL are open weight models from US and European developers, and Mistral Large 4 has the highest score of the four at 61.7%.
Excel modeling, tax and the Vals Finance Index
Vals AI runs two further finance benchmarks with tools and combines all three into the Vals Finance Index. The Excel Modeling Benchmark v1.1 has the model build LBO, DCF, M&A and operating models, comps spreads and dataroom summaries in Excel. Tax Agent Bench has research questions on US corporate tax; since September 2026, its primary metric counts an answer as correct only when every rubric check passes, with an adjustment for the share of cited sources that resolve to an authoritative document. The Vals Finance Index, first published in October 2026, weights the three benchmarks by each sub sector's share of US GDP.
| Benchmark | Highest closed source score | Highest open weight score | Difference | Last updated |
|---|---|---|---|---|
| Finance Agent v2.1 | 67.9% (Claude Opus 5.5) | 65.4% (DeepSeek V4.1 Flash) | 2.5 points | 5 Nov 2026 |
| Excel Modeling Benchmark v1.1 | 76.7% (Claude Fable 5.1) | 71.3% (Kimi K3) | 5.4 points | 5 Nov 2026 |
| Tax Agent Bench | 51.8% (Gemini 4 Carbon) | 44.4% (DeepSeek V4.1 Flash 1017) | 7.4 points | 5 Nov 2026 |
| Vals Finance Index | 67.1% (Gemini 4 Carbon) | 60.9% (DeepSeek V4.1 Flash 1017) | 6.2 points | 5 Nov 2026 |
On the Excel Modeling Benchmark, Kimi K3 ranks 13th of 39 models, followed among open weight models by Kimi K3.1 Pro at 70.0%, Step 5 Preview at 69.6% and DeepSeek V4.1 Flash 1017 at 69.2%. On Tax Agent Bench, DeepSeek V4.1 Flash 1017 ranks 7th of 82 models, above Muse Wonder Full at 43.0%, Claude Sonnet 5.5 at 42.0% and GPT 6.1 Sol at 21.8%. On the Vals Finance Index, DeepSeek V4.1 Flash 1017 ranks 9th of 37 models and GLM 5.3 10th at 60.5%.
Results by task
Finance Agent v2.1 scores each model on nine task categories and on All Pass, a strict metric that counts an answer as correct only when every check in it passes. The table compares the highest closed source score with the highest open weight score in each category.
| Category | Highest closed source score | Highest open weight score | Difference |
|---|---|---|---|
| Precedents | 45.6% (Claude Opus 5.5) | 49.3% (MiMo V2.6 Flash) | Open weight 3.7 points higher |
| Earnings Analysis | 90.3% (Muse Wonder) | 93.0% (DeepSeek V4.1 Flash 1017) | Open weight 2.7 points higher |
| Comparables | 62.7% (Gemini 4 Carbon) | 63.9% (Kimi K3.1 Pro) | Open weight 1.2 points higher |
| Overall | 67.9% (Claude Opus 5.5) | 65.4% (DeepSeek V4.1 Flash) | 2.5 points |
| General Quantitative Analysis | 87.8% (Claude Haiku 5.5) | 84.9% (Mistral Large 4) | 2.9 points |
| Market Analysis | 71.0% (Claude Sonnet 5.5) | 67.9% (GLM 5.3 Flash) | 3.1 points |
| General Qualitative Analysis | 79.2% (Gemini 4 Argon) | 75.3% (Muse Spark Open) | 3.9 points |
| Disclosure Analysis | 83.3% (Gemini 4 Argon) | 77.6% (Hy4 Preview) | 5.7 points |
| All Pass | 58.9% (Gemini 4 Argon) | 51.9% (GLM 5.3) | 7.0 points |
| Adjustments | 74.7% (Muse Wonder Full) | 67.5% (GLM 5.3) | 7.2 points |
| Financial Modeling | 57.0% (Claude Opus 5.5) | 46.8% (DeepSeek V4.1 Flash) | 10.2 points |
Categories led by open weight models
- Precedents: MiMo V2.6 Flash scores 49.3%, the highest score of all 37 models, against 45.6% for Claude Opus 5.5. This category covers M&A transaction multiples from S-4 filings and target financials with industry specific EBITDA normalization.
- Earnings Analysis: DeepSeek V4.1 Flash 1017 scores 93.0% and GLM 5.3 scores 92.4%, against 90.3% for Muse Wonder. This category compares reported results with consensus estimates, prior guidance and non GAAP reconciliations.
- Comparables: Kimi K3.1 Pro scores 63.9% against 62.7% for Gemini 4 Carbon. Inkling 2 Small, whose weights are announced, scores 65.0%. This category covers trading comps tables, EV multiples and enterprise value adjustments for off balance sheet items.
Largest gaps
- Financial Modeling: Claude Opus 5.5 scores 57.0% and DeepSeek V4.1 Flash scores 46.8%, a difference of 10.2 points. This category covers DCF, LBO and accretion and dilution models built from filing data.
- Adjustments: Muse Wonder Full scores 74.7% and GLM 5.3 scores 67.5%. This category bridges GAAP to non GAAP and pro forma figures across the income statement and cash flow statement.
- All Pass: Gemini 4 Argon answers 58.9% of questions with every check passed, and GLM 5.3 answers 51.9%.
Cost and speed
Cost per task
The table pairs open weight models with closed source models that reach a similar score on Finance Agent v2.1. Costs are API prices from the Vals AI run. Running an open weight model on the firm's own GPUs replaces the per token price with the cost of the hardware and its operation.
| Open weight model | Accuracy | Cost per task | Closed source model | Accuracy | Cost per task |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | 65.4% | $0.37 | Gemini 3.8 Flash | 65.8% | $1.99 |
| DeepSeek V4.1 Flash | 65.4% | $0.37 | Claude Haiku 5.5 | 65.4% | $1.64 |
| GLM 5.3 | 65.0% | $1.03 | Muse Wonder | 65.6% | $2.62 |
| MiMo V2.6 Pro | 64.3% | $0.18 | Claude Opus 5.5 | 67.9% | $8.91 |
| Mistral Large 4 | 61.7% | $0.35 | GPT 6.1 Sol | 61.8% | $1.63 |
| GLM 5.3 Flash | 60.6% | $0.07 | Grok 4.7 | 60.6% | $2.24 |
| Kimi K3 | 60.0% | $1.88 | GPT 6 Astra | 60.4% | $7.82 |
| DeepSeek V4 Pro 0813 | 57.1% | $0.82 | GPT 6 Luna | 56.8% | $0.12 |
MiMo V2.6 Flash has the lowest cost per task of the 20 highest scoring models at $0.07, followed by Inkling 2 Small at $0.18 and MiMo V2.6 Pro at $0.18. GLM 5.3 Flash reaches the same score as Grok 4.7 at 3% of the cost, and DeepSeek V4.1 Flash reaches the score of Claude Haiku 5.5 at 22% of the cost. MiMo V2.6 Pro is 3.6 points below Claude Opus 5.5 at 2% of the cost. The cost figures depend on the provider: Vals AI ran GLM 5.3 and GLM 5.3 Flash through Fireworks AI, Muse Spark Open through Together AI, Nemotron 3.5 Super VL through Baseten, and the DeepSeek, MiMo, Mistral, Hy4, Kimi, Step and MiniMax models through their developers' own APIs. Among closed source models, GPT 6 Luna reaches 56.8% at $0.12 per task, and Mercury 3 from Inception reaches 54.7% at $0.07.
Time per task
The top 5 open weight models take a median of 15.6 minutes per task on Finance Agent v2.1, and the top 5 closed source models take 29.7 minutes. Claude Opus 5.5 takes 33.0 minutes, Claude Sonnet 5.5 34.4 minutes and Gemini 4 Carbon 29.7 minutes, while Gemini 4 Argon takes 13.2 minutes and Muse Wonder 5.2 minutes. Among open weight models, GLM 5.3 Flash takes 4.9 minutes, Mistral Large 4 6.5 minutes, MiMo V2.6 Pro 11.4 minutes, DeepSeek V4.1 Flash 16.0 minutes and Kimi K3.1 Pro 35.0 minutes. Time per task includes every tool call and depends on the inference provider and the reasoning effort setting, which Vals AI set to max for the Claude, DeepSeek, GLM and Kimi models.
Data control and licensing
| Criterion | Open weight models | Closed source models |
|---|---|---|
| Where the model runs | On the firm's own servers, in its own cloud account, or through a hosted provider | On the developer's servers, accessed through an API |
| Where prompts and documents go | Stay inside the firm's infrastructure when self hosted | Sent to the developer under its API data terms |
| Model version | Fixed until the firm changes it | Updated or retired on the developer's schedule |
| Fine tuning on internal data | Possible on the published weights | Limited to the options the developer offers |
| Cost structure | Hardware and operations when self hosted, or per token through a provider | Per token or per request |
| License | MIT (DeepSeek, MiMo, GLM 5.3 Flash), Apache 2.0 (Hy4 Preview, Beam 1.0, Inkling) or a custom license with conditions such as revenue thresholds (Kimi, Muse Spark Open, Mistral Large 4, MiniMax, StepFun); see the license table in the open source models ranking | Developer's commercial terms of service |
| Accuracy on Finance Agent v2.1 | Highest score 65.4% | Highest score 67.9% |
FAQ
Are closed source models better than open source models for financial analysis?
Closed source models have the higher overall scores on the Vals AI finance benchmarks as of November 2026. The difference is 2.5 points on Finance Agent v2.1, 5.4 points on the Excel Modeling Benchmark v1.1, 7.4 points on Tax Agent Bench and 6.2 points on the Vals Finance Index. In three Finance Agent v2.1 categories, Precedents, Earnings Analysis and Comparables, an open weight model has the highest score.
What is the best open source model compared with GPT, Claude and Gemini for finance?
DeepSeek V4.1 Flash has the highest open weight score on Finance Agent v2.1 at 65.4%, followed by GLM 5.3 from Z.ai at 65.0% and MiMo V2.6 Pro from Xiaomi at 64.3%. Claude Opus 5.5 scores 67.9%, Gemini 4 Argon 67.7% and GPT 6.1 Sol 61.8%. DeepSeek V4.1 Flash ranks 9th of 37 models and costs $0.37 per task. On Tax Agent Bench and the Vals Finance Index, DeepSeek V4.1 Flash 1017 has the highest open weight score.
Are open source models cheaper than closed source models for financial analysis?
At similar accuracy, most open weight models cost less per task through hosted APIs. DeepSeek V4.1 Flash costs $0.37 per task at 65.4%, and Claude Haiku 5.5 costs $1.64 at the same score. GLM 5.3 Flash costs $0.07 per task and Grok 4.7 $2.24, both at 60.6%. The median cost per task of the top 5 models is $0.37 for open weight models and $6.95 for closed source models.
Are open source models slower than closed source models?
On Finance Agent v2.1, the top 5 open weight models take a median of 15.6 minutes per task and the top 5 closed source models take 29.7 minutes. In September 2026 the order was reversed, at 14.8 and 3.3 minutes. Time per task depends on the number of tool calls, the reasoning effort setting and the speed of the inference provider. Claude Opus 5.5 and Claude Sonnet 5.5 run at max reasoning effort and take more than 30 minutes per task.
Which financial tasks show the largest difference between open and closed source models?
Financial Modeling shows the largest difference on Finance Agent v2.1 at 10.2 points, followed by Adjustments at 7.2 points and the strict All Pass metric at 7.0 points. On Tax Agent Bench the difference is 7.4 points, and on the Excel Modeling Benchmark, which tests model building in Excel, it is 5.4 points.
How do the November 2026 results compare with September 2026?
Vals AI replaced Finance Agent v2 with Finance Agent v2.1 and the Excel Modeling Benchmark with v1.1, and changed the primary metric of Tax Agent Bench, so the November scores are on different question sets and grading rules. On Finance Agent v2 in September, the difference between the highest closed source and open weight score was 3.6 points (Gemini 3.8 Flash at 61.4% and GLM 5.3 Flash at 57.9%). On Finance Agent v2.1 in November it is 2.5 points.
Related reading
- Best open source models for financial analysis: ranking, licenses and hardware requirements
- How to access financial statements with ChatGPT or Claude
- How to analyze stocks with AI
- Financial Analysis Prompt Library
- Vals AI Finance Agent v2.1
- Vals AI Excel Modeling Benchmark v1.1
- Vals AI Tax Agent Bench
- Vals Finance Index
Last updated: 6 November 2026. Benchmark scores, prices and model versions change with each release. Educational content for general information about AI models.